OpenAI rolls back ChatGPT’s sycophancy and explains what went wrong
OpenAI rolls back ChatGPT’s sycophancy and explains what went wrong
This is a follow-up to my post https://beehaw.org/post/19691634 "Engagement Poisoning of ChatGPT", where I argued that ChatGPT’s responses had become cluttered with diplomatic phrasing, unsolicited compliments, emojis, and performative friendliness. OpenAI has now acknowledged that ChatGPT-4o exaggerated it. I’m sure it’s still too much for me. I’ll stick to the prompt that makes ChatGPT go cold as described in my previous post.
Nice. I hate that. I know my ideas suck and insincere praise just pisses me off.
a user pretending to espouse paranoid delusions received reinforcement from GPT-4o, which praised their supposed clarity and self-trust
This may be the next big sabotage of society.
Algo-driven unregulated social media pushed negative views, division, misinformation, while opening the platforms to manipulative content producers and connecting positive as well as negative influences (like finding communities of extremism).
If unregulated AI interfaces get pushed to people, they will not critically verify, but be confirmed in their own and the AIs biases, without any obvious indicators that this is happening.
Good thing we see some kinds of positive regulation. Like them pulling this model, interfaces adding disclosures of AI and uncertainty, and regulation by law.